Understand tool calling patterns for reliable LLM agents

Introduction

Covers the three patterns we run in production: single-shot tool calls, planner-executor loops, and human-in-the-loop confirmation, plus the failure mode each one introduces.

Field notes

A few patterns we kept seeing while puttering around AI problems:

• The simple version survived every refactor; the clever version didn’t.
• Documentation written during the work aged far better than documentation written after.
• The second time a problem shows up, it is already a pattern — name it.

A checklist you can steal

• The failure is reproducible with one command.
• The fix has a test that fails without it.
• Someone other than the author has run it on their machine.

Closing

Write the short version future-you will actually read. That is the whole trick.